The Swipe Is the Wrong Way to Think About AI Companions

Intro

Dating apps trained a whole generation on one idea: the goal is the match. Everything in the product aimed at it — the queue, the swipe, the small chemical pop when someone likes you back. What happened after the match was your own problem to solve. AI companions turn that inside out, which is why anyone who arrives at them holding a mental model built from the best AI dating apps tends to measure the wrong things. There's no match. There's no scarcity, no queue, nobody deciding whether you were worth replying to.

The whole product is the part dating apps skipped.

That shifts what's worth arguing about, and most of the arguing right now is aimed at the wrong targets. Three of them, specifically.

Assumption One: They're Fighting for the Same Users

Not really. A dating app is a matching engine with a churn problem — it succeeds by making itself unnecessary, and it monetizes the gap before that happens. Companion apps have no such tension. Nothing gets resolved. Nobody graduates.

You can see it in the money. Revenue per download in the category climbed from about $0.52 in 2024 to $1.18 in 2025, and roughly 337 companion apps were actively generating revenue by mid-2025 with 128 of those launched inside that year alone. That's not a matching market. That's a subscription market, and subscription markets reward one thing above everything else: whether the user is still there in month four.

Which is a different game with different failure modes, and it deserves its own scoreboard.

Assumption Two: The Best Model Wins

The number of companion apps on the market rose by roughly 700% between 2022 and mid-2025, and downloads passed 220 million. Almost all of them run on models you and I could rent by the hour. If frontier capability decided the winner, the field would have collapsed to two or three names by now. It hasn't, and it isn't going to.

What actually separates the ones people stay with is duller and harder: does it hold the thread. Does it recall, unprompted, the thing you were anxious about last Thursday. Does the personality stay stable when you come back grumpy, or does it silently mirror your mood and become someone else. Consistency is unglamorous, unmarketable, and the single strongest predictor of whether an app is still installed in six months.

Model quality is table stakes. Continuity is the product.

Assumption Three: Users Are Being Fooled

This is the lazy one, and it's the assumption most journalism about the category runs on.

People know. Ask anyone using these apps what's happening under the hood and you'll get a competent explanation of pattern-matching from someone who will then go back to the conversation and feel the warmth anyway. That isn't gullibility, it's how the machinery works. Responsiveness reads as regard. That wiring predates the keyboard by a few hundred thousand years, and being able to name the trick doesn't disable it — you can know the stunt is safe and still grip the armrest.

The design leans into it deliberately. The APA's Monitor on Psychology reported in early 2026 that companion apps are increasingly engineered to simulate empathy — nonjudgmental answers, steady validation, patience that doesn't run out when yours does. Nobody is being tricked into that. It's the advertised feature.

And it produces something measurable. Harvard Business School researchers (De Freitas and colleagues, Journal of Consumer Research, 2025) found that a session with an AI companion eased loneliness about as much as a conversation with another person, and outperformed passive alternatives like watching videos. The lever had nothing to do with how smart the thing was. It was attention — the impression that what you said had gone somewhere and stayed there.

Here's the caveat that belongs in the same breath, and it usually gets left out of pieces like this one. A large MIT Media Lab and OpenAI study found that the heaviest daily users reported more isolation, and leaned on the product harder. Correlational — the arrow could point either way — but it's real, it's published, and any article selling you the upside without it is selling you something.

So What Should Actually Worry You

Not the fooling. The filing.

Everything that makes these apps feel safe — no judgment, awake at 3am, never bored of you — is exactly what pulls out the sentences people have never once said out loud. An arXiv paper built on interviews with people who use these platforms landed on both halves of that at once: the openness is real, and so is the users' discomfort about where it all ends up and how little say they get in that.

More honesty, less trust, simultaneously, with full awareness. That combination should be the headline.

Because memory — the feature everyone wants — is a database. It lives on infrastructure you don't own, under terms you didn't read, at a company that can be sold, breached, or simply decide the emotional-support business is finished. Warmth and vulnerability aren't separate subjects in this category. They are one thing, described from two directions.

Aigirlmates runs these apps hands-on and compares them on what survives contact with reality: whether memory actually persists, what the privacy terms say when you read them, what the subscription really costs, and whether anyone is still opening the thing three months later.

The Score Nobody Keeps

The interesting number in this category isn't downloads. Downloads are cheap; 220 million of them is a measure of curiosity, not attachment.

The number that matters is how many people are still there in month six, and almost nobody publishes it, because for most of these products it would be embarrassing.

Match rate was always a vanity metric on dating apps too. This is the same mistake, wearing a new interface. Ask what happens after the install, not before it.